Generate a unique identifier like ALPHA-3F8D-2C91 — short enough to fit naturally in a sentence, long enough to never collide with normal text

Strip body fields the target model is known to reject. Tolerates the common dot-vs-hyphen typo (e.g. user types `claude-opus-4.7`) by normaliz

The retrieval prompt appended after the haystack. Uses an explicit "report each on its own line" structure so we can parse model output relia

Run official OpenAI probes and return a JSON-serializable report.

POST /v1/messages/count_tokens. Anthropic's count endpoint estimates input tokens for the same request shape used by Messages. Relays

Extract stable protocol fingerprints from one raw response.

Bidirectional prefix match (alias <-> snapshot tolerance), with dot-vs-hyphen normalization.

Validate a raw OpenAI response payload against a wire API template.

Drop-in for _run_anthropic/_run_openai/_run_gemini that captures the passed-in ExecutionConfig and returns a minimal valid outcome shape.

Resolve the originating client IP. uvicorn is started with --proxy-headers --forwarded-allow-ips=* so by the time FastAPI sees the request, r

completion_tokens must not exceed the request cap. For thinking models (e.g. Gemini 3) the field includes reasoning tokens and can fill the e

Strip protocol + path, keep host. example: https://api.x.com/v1 → api.x.com.

Minimal SSE parser. Each event is a block of lines terminated by blank line. We only care about `event:` and `data:` lines. `data:` may span mult

Helper: spec is list of (domain, [score1, score2, ...]).

Return whether stream usage looks compatible with a non-stream probe. For reasoning-capable models, the same short prompt can spend a different

Date string from a template's mtime, or '' if missing/unreadable.

Return False to skip this detector for the given model.

Build low-cost probes that exercise official OpenAI wire features.

Test helper.

Suggested base URL placeholder for the Gemini submission form.

Streaming variant. Note: streaming requests still respect the semaphore, but retry logic only applies to the initial connect — once e

Test helper — wipe all buckets.

Update per-stream local usage from an event. message_start carries true input_tokens; message_delta carries cumulative output_tokens